fix(io): return an error for short reads

[lance-gatekeeper]

Summary

• validate physical read lengths before reconstructing coalesced requests
• return an I/O error with the file, byte range, actual length, and expected length
• add a regression test for an empty short read

Root cause

FileScheduler assumed every reader result exactly matched its requested physical byte range. When a reader returned fewer bytes, request reconstruction called Bytes::slice with the expected end offset and panicked before reporting which read was incomplete.

Validation

• cargo fmt --all
• cargo test -p lance-io test_short_read_returns_io_error
• cargo clippy --all --tests --benches -- -D warnings
• cargo test -p lance-io (190 tests passed; 8 unrelated io_uring tests could not initialize in this container because IoUring::new returned EPERM)

Fixes #5674

[lance-gatekeeper]

Gate recommendation: approve. The shared post-read validation enforces the exact physical-range contract before coalesced and split reconstruction, converting short-read panics into contextual I/O errors while covering both scheduler backends and custom readers.

[westonpace]

It might be premature optimization but it would be nice if we could avoid the path clone on every IOP. If we move the validation lower we might be able to avoid this?

[lance-gatekeeper]

Addressed in 177c716: moved short-read validation into the physical scheduler paths, avoiding the per-I/O path clone while preserving eager lite submission and contextual I/O errors.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The revision enforces exact physical-read lengths inside both scheduler backends, eliminating the eager path clone while preserving typed errors and the existing coalescing, ordering, and backpressure contracts.

[Xuanwo]

Nice change, thank you!